International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 06 Issue: 07 | July 2019
p-ISSN: 2395-0072
www.irjet.net
THE SENTIMENTAL ANALYSIS ON PRODUCT REVIEWS OF AMAZON
DATA USING THE HYBRID APPROACH
Kajal 1*, Prince Verma2
1M.Tech
Scholar, Dept. of Computer Science Engineering, CT Institute of Engineering Management & Technology,
Jalandhar, India
2H.O.D & Assistant Professor, Dept. of Computer Science Engineering, CT Institute of Engineering Management &
Technology, Jalandhar, India
----------------------------------------------------------------------***--------------------------------------------------------------------Abstract - The term sentiment specifies to the viewpoint or
speculation of the person towards some particular domain like
the topic, product information, etc. The sentiments are
analyzed by the user towards any domain or elements by using
this SA approach. Web or internet or social network is the
best-known source to assemble sentiment information. Among
popular social network portals or sites, Twitter has been the
point of attraction to lot of researchers. Twitter is a big
platform in which user post their messages or point of view
related to any product or any other thing are said to tweets. In
this paper, the study work is based on the sentiment analysis
(SA) of product reviews of Amazon data by using the Twitter
API .In the existing work, the SVM technique is applied that
contains some issues and these issues can be resolved by
applying the KNN technique for the sentiment analysis. We
used six different training-testing compositions for
performance analysis. Moreover, we present the performance
contrast of the discussed techniques based on our recognized
parameters.
Key Words: KNN (K-Nearest Neighbor), SA (Sentiment
Analysis), Sentiments, SVM (Support vector machine),
Twitter.
I.
INTRODUCTION
Twitter moreover another Social media (like facebook,
WhatsApp etc), Social Networks, online posts and
microblogging sites on the web became closely the largest
destinations on the web for communication in between
people or any other party to show their thoughts related to
any products [1]-[2] or movies [3] allotment their regular
understanding and convey their point of view or thoughts
about actual time and upcoming events like political
elections or sports[4] etc. The mining of such type of
thoughtful data that can be extracted from social media is
highly useful to conclude various fields [5]. Inappropriate,
the growth of social media (SM) such as Twitter, Facebook,
Snap-chat, Instagram, Wechat and also online reviews has
powered interest in sentiment analysis [6].
"neutral" is completed through sentiment analysis way.
Since getting any kind of products all user gain the
knowledge about them through sentiment analysis way.
Based on user’s requirement the products or services can be
served by the associations or organizations to their buyers
or purchasers. The available accurate data is processed
figure out or examined by the technique like textual
information retrieval. The subjective attributes can be
indicated by some textual contents or facts. The base of
sentiment analysis is organized by the essence which mainly
involves thoughts, views, opinions, and sentiments [7].
In broad, there are three levels where thoughts mining based
on which sentiment analysis (SA) can be executed. They are
given below:
a. Document-level sentiment analysis: In Document level, SA
classifies the whole document and shows the polarity like
positive or negative or neutral for any kind of service or
product. It is the simplest type of sentiment analysis. Mostly
under the document-level sentiment analysis [8], we can use
two types of approaches that are shown below:
[1] Supervised learning
[2] Unsupervised learning
In the supervised learning approach that considers there is a
definite no of groups or classes in which the whole
documents should be classified and for individual class
training data is accessible. In the Unsupervised learning
approach, the analysis is with-in document conclusive the
semantic direction thus of unique phrases. Mostly we can
select the phrases by two main methods that are given
below:
[1] A set of predefined part of speech
[2] A lexicon of sentient words [8]-[9].
In sentiment analysis, the Tweets, text, and speech that show
opinions, thoughts, views, and emotions related to any
product or service are extracted from the database sources.
Arrangement of thoughts amid "positive", "negative" and
b. Sentence-level sentiment analysis: To determine a
particular sentence show a positive, negative or neutral
thoughts or feelings for any kind of product or service, we
can use Sentence-level sentiment analysis. This level of
analysis is executed by two tasks that are subjective and
objective. Among which the Subjective sentence is used for
the classification is executed and the true information
expressed by objective sentences [10]-[11].
© 2019, IRJET
ISO 9001:2008 Certified Journal
|
Impact Factor value: 7.34
|
|
Page 1539
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 06 Issue: 07 | July 2019
p-ISSN: 2395-0072
www.irjet.net
The concluded results indicated that in the future, sentiment
analysis will be done on other social networking sites also.
Document-level
sentiment analysis
Sentiment
Analysis
Level
Sentence-level
sentiment analysis
Aspect-based sentiment
analysis
Fig 1: Sentiment Analysis Level
c. Aspect-based sentiment analysis: This approach work on a
single entity.The research problem which targets on the
recognition and extraction of every sentiment expressions
within a provided document [11].
This paper contains 5 sections that describe the sentiment
analysis of Twitter data. Section I contains the Introduction
about the sentiment analysis, its levels. Section II presents
various literature surveys in the field of sentiment or opinion
mining. Section III represented the sentiment analysis
research methodology and Section IV elaborates
experimental results and discussion of work. The Last
section V defines the conclusion and future scope of
sentiment analysis.
II.
LITERATURE REVIEW
From the last set of years, many articles, papers, and books
have been written on sentimental analysis. At a similar time,
some researcher's center of attention more on exact burden
like finding the subjectivity expression, subjectivity clues,
topics, and sentiments of words and extracting sources of
opinions, while others target is on assigning sentiments to
the whole document. All analysts of sentiment analysis have
become distinct approaches to naturally predict the
expression, sentiments of words or a document. The data set
for sentimental analysis considered are a movie, product
review or social media data from the source of the internet.
Because of the importance of the sentiments analysis,
researchers have developed different techniques to enhance
the efficiency of the analysis. They use pattern-based
approaches, NLP (Natural Language Processing), lexicon
based analysis as well as machine learning techniques, etc.
Huma Parveen, et.al (2016) proposed the Hadoop
Framework that was used for processing the data set of
movie accessible on the twitter in a variety of feedback,
reviews, and comments [12]. The outcomes of SA on twitter
data displayed in the type of different sections similar to
positive, negative and neutral sentiments. The preprocessing of data was done to remove noise. This type of
analysis would help any organization to increase their
business productivity. This method also provided the fast
downloading approach for proficient twitter trend analysis.
© 2019, IRJET
|
Impact Factor value: 7.34
|
Ankit, et.al (2018) proposed that Ensemble Classifier was
used to improve the precision and performance of the
sentiment classification technique [13]. This classifier
combined the base learning classifier to form a solitary
classifier. The obtained results indicated that the ensemble
classifier gave better performance than a stand-alone
classifier. In sentiment classification technique, the role of
data pre-processing and feature representation was also
explored. The ensemble classifier system was developed by
using various base learners like Naïve Bayes classifier, SVMs,
random forest classifier, and logistic regression. This
approach was very helpful for companies to monitor
consumer opinions regarding their products. In the future,
the main center of attention will be on the study of natural
tweets since the different tweet sentiments.
Sushree Das, et.al (2018) stated that a stream-based
setting by using the incremental active learning approach
gave the capability to the algorithm for choosing new
training data from a data stream for hand-labeling [14].
Stream-based active learning in the financial domain could
be helpful to both sentiment analysis and the active learning
research area. With the use of RNNs Long Short –Term
Memory, this experiment also proved helpful for feasibility
study through batch processing. To analyze the sentiments
and recent stock trends, a hybrid model could also be
developed. This model would improve the reliability of the
prediction. In the future for analyzing the stock data,
addition also applies the machine learning algorithms. Some
additional methods of data ingestion like data ingestion
through the Apache Flume or NodelJS can also be used in the
future.
Symeon Symeonidis, et.al (2018) proposed that various
pre-processing techniques evaluated on their ensuring the
accuracy of the classification and the number of features
they developed [15]. The obtained results indicated that
some methods like lemmatization and removing numbers
also replacing contractions improved precision while other
techniques like removing punctuations did not. To explore
the interactions between the methods when they were
employed in a pipeline manner, an ablation and combination
study was done. The outcomes of these techniques indicated
the importance of techniques like replacing numbers and
replacing repetitions of punctuations.
Neethu M S, et.al (2013), in this paper, they analyze the
twitter data related to Electronic products with Machine
Learning approach [16]. They existent a new Feature-Vector
for classification of the tweets and extricate peoples'
opinions about Electronic products.Thus created a FeatureVector from 8 relevant features. Naïve-Bayes and SVM
classifiers are implemented using built-in functions of
Matlab. Max-Entropy classifier is implemented using
ISO 9001:2008 Certified Journal
|
Page 1540
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 06 Issue: 07 | July 2019
p-ISSN: 2395-0072
www.irjet.net
Maximum-Entropy software. All the used classifiers have
almost equal performance.
the system can be determined based on an output that is
generated by the system.
Rasika Wagh, et.al (2018) reviewed that the sentiment
analysis involved the area of data mining and NLP [17]. The
sentiment analysis research of Twitter data could be done in
many ways. Various types and techniques of sentiment
analysis were used to done extraction of sentiments from
tweets after which a comparative study is done. In classify to
construct opinion on sentiment, the analyzation of twitter
data was done from a different point of view. The study of
the literature indicated that when the semantic analysis
Word-Net was followed up by the machine learning
techniques like SVM, Naïve-Bayes and maximum entropy
then the accuracy improved. The accuracy could also be
increased by up to 4-5% by using the hybrid approach.
B. Preprocessing the data
M.Trupthi, et.al (2017) proposed an interactive automatic
system that used Hadoop to analyze the sentiment of the
tweets posted in social media [18]. In the process of
Sentiment analysis, genuine tweets were used. The main
motive of research was to perform real-time sentiment
analysis on tweeter data and to give time-based prediction to
the user. The main features of the system were a training
module which was done by using Hadoop and Map-Reduce.
The classification method was based on Naïve Bayes, TimeVariant Analytics and the Continuous Learning system. The
proposed system had some limitations like the use of Unigram Naïve, adaptability to only English Language and the
lack of actual intended meaning. But these problems could
be removed by making certain changes like the use of ngram classification instead of Uni-gram, to build more
adaptive language systems, etc. In the future, this system
will be helpful to people and industries based on sentiment
analysis like Sales Marketing, Product Marketing, etc.
III.
SENTIMENTAL
METHODOLOGY
ANALYSIS
In this segment, the input data is pre-processed before
extracting the features. Dataset used in EXCEL or CSV file is
used and comments are expressed in ENGLISH language.
Stemming [22], error correction and stop word removal are
the three main preprocessing techniques which are
performed here [20].
Twitter
API
Standard Product dataset
(Excel Sheet)
Preprocessing the data
Lexical Analysis of sentences
N-Gram for Extraction of features
Define positive, negative and neutral words
Senti-word net Usage
RESEARCH
WDE-K-Nearest Neighbor Classifier
This section presents the steps of the sentiment analysis of
twitter data shown in fig 2. The method classifies a Twitter
data set of English language into positive or negative or
neutral sentiments.
Output: Number of positive, negative and
neutral
A. Standard Product Dataset
We gathered our dataset by consulting the Twitter API. Two
types of datasets are generated manually here amongst
which one is used for training and another is used for testing.
X: Y is the relation present within the training set. Here the X
is represented by the score of probable opinion word and Y
is the representation of whether the score is positive or
negative. By gathering reviews from the e-commerce sites,
the testing set is generated. A review of whether the testing
set is positive or negative is manually tagged. The reviews
will be separated based on positive and negative sentiments
they include once the training is completed after that the
system is tested, with the help of reviews. The accuracy of
© 2019, IRJET
|
Impact Factor value: 7.34
|
Stop
Fig 2: Steps of Proposed Sentimental Analysis Flowchart
C. Lexical Analysis of Sentences
In the dataset, Sentence which includes either a positive or
negative sentiments. To minimize the complete size of the
review, such sentences can be removed which might not
include any sentiments within them. The regular expressions
involved within python do not recognize these sentences of
questions [19].
ISO 9001:2008 Certified Journal
|
Page 1541
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 06 Issue: 07 | July 2019
p-ISSN: 2395-0072
www.irjet.net
D. N-Gram for Extraction of Features
The main problem arises inside the SA though extractive the
features from data. The N-gram algorithm is applied which
can extract the features and also post tag the sentences.
E. Define Positive, Negative and Neutral Words
The grammatical dependencies present the words in the
sentences will be gathered by the parser and given as output
[18].To identify the opinion word for features that have been
gathered from the last step, the dependencies have to be
looked upon in further steps [19]. There is also a need to
contain the direct dependencies and transitive dependencies
within this step [20].
ratio )i.e. 50 percent of the data was used for training and 50
percent of the data was used for testing, 85.92%(in the case
of 60-40) i.e. 60 percent of the data was used for training and
40 percent of the data was used for testing, 91.08%(in case
of 70-30 ) i.e. 70 percent of the data was used for training
and 30 percent of the data was used for testing, 91.15 %( in
case of 80-20) i.e. 80 percent of the data was used for
training and 20 percent of the data was used for testing and
96.43( in case of 90-10) i.e. 90 percent of the data was used
for training and 10 percent of the data was used for testing.
This experimental result shows that the KNN approach
achieved the highest accuracy for the classes like positive,
negative and neutral than the SVM approach, as elaborated
in Chart 1.
Table 1: Accuracy Analysis
F .SentiWordNet Usage
Sentiwordnet is generated especially within the opinion
mining applications. There are 3 main relevant polarities
present for each word within the Sentiwordnet which are
positivity, negativity and neutral [17]-[20].
G. WDE-K-Nearest Neighbor Classifier
In organize to use a classifier within this approach, WDEKNN is preferred. Since sentiment analysis (SA) is a binary
classification and there are huge datasets that can be
executed, WDE-KNN is chosen here. The generated training
set is useful for training the classifier here. There is X: Y
relation provided within the training set. In which the x is
represented by the score of an opinion word and it is
represented by score whether the word is positive or
negative [17]-[20]-[21]. A score of the view word related to a
feature within the review is specified as input to the KNN
classifier.
Training ,Test
Ratio
SVM classifier
(%)
KNN classifier (%)
40-60
80.63
84.63
50-50
81.17
85.32
60-40
82.59
85.95
70-30
84.17
91.08
80-20
85.88
91.15
90-10
85.92
96.43
H. Output
The output comes in the form of extraction of features wise
opinions i.e. Number of positive, negative and neutral. The
accuracy of the system can be decided on the base of output.
IV.
EXPERIMENT RESULT & DISCUSSION
In this research work is related to sentiment analysis of
twitter dataset. This section presents the environment
setups; the data is collected over the twitter by using Twitter
API and the conducted experiments for studying the method
performance.
1). Accuracy: For Accuracy analysis, Table 1 shows the
overall result and the graphical representation
demonstrated in Chart 1. The classifiers used were SVM and
KNN for classifications. The classifiers were trained and
tested several times on the same dataset. KNN classifier
achieved high accuracy 84.63 %( in case of 40-60 ratio) i.e.
40 percent of the data was used for training and 60 percent
of the data was used for testing , 85.32% (in case of 50-50
© 2019, IRJET
|
Impact Factor value: 7.34
|
Chart 1: Accuracy Analysis
2).Execution Time: For Execution Time, Table 2 shows that
overall results for the different number of training and test
ratio. The graphical representation of all these results
presented in Chart 2. In case of (40-60 ratio) i.e. 40 percent
of the data was used for training and 60 percent of the data
was used for testing and the experiment result show that
KNN approach required only 3.08625 seconds which is
comparatively low as compared to the SVM approach. In case
of (50-50 ratio) i.e. 50 percent of the data was used for
training and 50 percent of the data was used for testing and
the experiment result show that KNN approach required
only 3.49785 seconds which is comparatively low as
ISO 9001:2008 Certified Journal
|
Page 1542
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 06 Issue: 07 | July 2019
p-ISSN: 2395-0072
www.irjet.net
compared to the SVM approach. In the case of (60-40 ratio)
i.e. 60 percent of the data was used for training and 40
percent of the data was used for testing and the experiment
results prove that the KNN approach required only 3.27577
seconds which is comparatively low as compared to the
approach SVM.
In the case of (70-30 ratio) i.e. 70 percent of the data was
used for training and 30 percent of the data was used for
testing and the experiment results prove that the KNN
approach required only 3.83721 seconds which is
comparatively low as compared to the SVM approach. In the
case of (80-20 ratio) i.e. 80 percent of the data was used for
training and 20 percent of the data was used for testing and
the experiment results prove that the KNN approach
required only 3.44296 seconds which is comparatively low
as compared to the SVM approach. In the case of (90-10
ratio) i.e. 90 percent of the data was used for training and 10
percent of the data was used for testing and the experiment
results prove that the KNN approach required only 3.37808
seconds which is comparatively low as compared to the SVM
approach.
Table 2: Execution Time
Training ,Test
Ratio
SVM classifier
(Time in
seconds)
KNN classifier
(Time in
seconds)
40-60
7.07257
3.08625
50-50
8.27024
3.49785
60-40
9.22069
3.27577
70-30
9.76641
3.83721
80-20
10.5381
3.44296
90-10
11.26555
3.37808
This experiment result proved that the KNN approach
required low time (in seconds) for execution as compared to
the SVM approach, as presented in Chart 2.
V.
CONCLUSIONS AND FUTURE SCOPE
Sentiment analysis of tweets is an important area of
research. The sentiment analysis of Twitter data, we can
analyze user behavior. Through classification and analysis of
sentiments on Twitter, we can get perceptive of people's
attitudes about the product. Firstly N-gram approach is
applied for sentiment analysis of Twitter data. By using Ngram we can extract the features then analyzed the input
data. Furthermore, We applying a classification technique for
data classification. In this work, we can study and
comparative analysis the techniques used are SVM and KNN.
The classification of KNN performs better as compared to the
classification of SVM because KNN uses multiple hyperplanes for data classification. The SVM and KNN techniques
are implemented in Python and Experiment results analysis
shows that the KNN approach achieved a high accuracy score
over the SVM approach. The execution time of the KNN
approach is low as compared to SVM. In the future, we would
similar to analyze and compare sentiment analysis with
other domains.
REFERENCES
[1] B. O'Connor, R. Balasubramanyan, BR. Routledge, and N.
Smith, ''from tweets to polls: Linking text sentiment to public
opinion time series,'' in Proc. International. AAAI Confer.
Weblogs Social Media, pp. 26-33, May 2010.
[2] M. Anthony Cabanlit and K. Junshean, '' Optimizing NGram based text feature selection in sentiment analysis for
commercial products in Twitter through polarity lexicons,''
in Proc. 5th Int. Conference. Inform, Intell. Syst. Application.,
pp. 94-97, Jul. 2014.
[3] Umesh Rao Hodeghatta, ''Sentiment analysis of
Hollywood movies on Twitter,'' in Proc. IEEE/ ACM
ASONAM, pp. 1401-1404, Aug. 2013.
[4] Juan M. Soler, Fernando Cuartero, and Manuel Roblizo,
‘‘Twitter as a tool for predicting elections results,’’ in Proc.
IEEE/ACM Int. Conf. Adv. Social Netw. Anal. Mining
(ASONAM), 2012, pp. 1194–1200.
[5] R. Barahate Sachin and M. Shelake Vijay, "A Survey and
Future Vision of Data mining in Educational Field", in Proc.
2nd Int. Conf. on Advanced Computing & Communication
Technology, pp. 96-100, 2012.
[6] Liu B. “Sentiment analysis and opinion mining.” San
Rafael, Computer Application: Morgan & Claypool, 2012.
[7] R. Parikh and M.Movassate, "Sentiment Analysis of the
User- Generated Twitter Update using Various Classification
Technique", Final Report CS224N , 2009.
Chart 2: Execution Time Analysis
© 2019, IRJET
|
Impact Factor value: 7.34
|
ISO 9001:2008 Certified Journal
|
Page 1543
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 06 Issue: 07 | July 2019
p-ISSN: 2395-0072
www.irjet.net
[8] R. Feldman, "Techniques and Applications for Sentiment
Analysis", Communications of the Association for Computing
Machinery, pp 8289, Volume. 56, Issue-4, 2013 .
Inter. Journal of Innovative Tech. and Exploring Eng.
(IJITEE), Vol-8 , Iss-8, 2019
[9]E.Divya,"Real Time Sentiment Classification Using
Unsupervised Reviews" Inter. Jour. of Sci. & Engi..Research
(IJSER), Vol 5(3), March-2014.
[10] R. Kaur, P. Verma "Sentiment Analysis of Movie
Reviews: A Study of Machine Learning Algorithm with
Various Feature Selection Methods", (IJCSE), Vol-5, Iss-9,
2017.
[11] N. Nehra, "A Survey on Sentiment Analysis Of Movie
Reviews", (IJIRT) , pp 36-40,Volume1, Iss.7, 2014.
[12] H. Parveen, P. S. Pandey, "Sentiment Analysis on Twitter
Data-set using NB Algorithm", Senti WordNet Dictnary,2016.
[13] Ankit, Nabizath Saleena, "An Ensemble Classification
System for Twitter Sentiment Analysis" , Int. Conf. on Comp.
Int. and Data Sci. (ICCIDS ),2018.
[14] S. Das, R. Kumar Behera, M. Kumar, S. Kumar Rath,
"Real Time Sentiment Analysis of Twitter Streaming Data
For Stock Prediction", Int. Conf. on Comp. Int. and Data Sci. ,
2018.
[15] Sy. Symeonidis, Di. Effrosynidis, Avi. Arampatzis, "A
comparative evaluation of pre-processing techniques and
their interactions for twitter SA", Expert Sys. With App.
(ESA), 2018.
[16] Neethu, R. Rajasree and M. S., "Sentiment analysis in
twitter using machine learning techniques." Comp., Comm.
and Net. Technologies (ICCCNT), IV Int. Conf. on IEEE, 2013.
[17] R. Wagh, P. Punde, "Survey on Sentiment Analysis using
Twitter Dataset", Proc. of the 2nd Inter. conf. on Electronics,
Communication and Aerospace Tech. (ICECA ) ,2018.
[18] M.Trupthi, G.Narasimha, S. Pabboju, "Sentiment
Analysis on Twitter using Streaming API", 7th Inter. Advance
Comp. Confer., IEEE, 2017.
[19] Z. Rezaei, M. Jalali, “Sentiment Analysis on Twitter using
McDiarmid Tree Algorithm", 7th Inter. Confer. on Computer
and Knowledge Eng. (ICCKE ), October 26-27 ,2017.
[20] Kajal, Prince Verma "Hybrid approach for sentimental
analysis of twitter data", Inter. Jour. of Comp. Sci. and Eng..,
Volume-7 (6), Jun 2019.
[21] Martín-V. M T, Rushdi S. M, Urena-Lopez L A, MontejoRaez A, "Experiments with SVM to classify opinions in
different domains", Expert Sys. With Applications (ESA),
pp:14799- 14804, 2011.
[22] H. Kaur, Prabhjeet Kaur, "Dimensionality Reduction In
Sentiment Analysis Using Colony-Support Vector Machine",
© 2019, IRJET
|
Impact Factor value: 7.34
|
ISO 9001:2008 Certified Journal
|
Page 1544